A Ten-thousand-word Interview with Ultraman: Unveiling the Hardships Behind GPT-5 and Declaring the Eve of Superintelligence Has Arrived
Key points:
The programming and creative capabilities of GPT - 5 have leaped. It can complete complex programming in 7 seconds, and its writing is more natural. It has become a landmark tool for "instantly creating customized software".
Sam Altman predicts that a generally recognized major scientific breakthrough may occur within two years. Currently, what is needed is to improve the cognitive ability of the model.
The development of AI is restricted by four core bottlenecks: computing power, data, algorithm design, and clear product goals, which jointly determine the evolution speed and implementation value of AI.
For those willing to invest "cognitive load time", AI is a multiplier. In the future, a single person may start a billion - dollar company, and the thresholds for creativity and execution will be significantly lowered.
On August 8th, after the release of GPT - 5, Sam Altman, the CEO of OpenAI, was a guest on the "Huge Conversations" interview and received an exclusive interview from well - known YouTube anchor Cleo Abram. They delved into the future development of artificial intelligence (AI) and its profound impact on society. While recalling the difficult days of programming the "Snake" game on a calculator when he was 11 years old, he demonstrated how GPT - 5 could generate the same game in 7 seconds. This was not only a technological demonstration but also seemed to announce that the eve of super - intelligence had arrived.
In this conversation of over ten thousand words, he first talked about the detours and surprises in the R & D behind GPT - 5, candidly admitted the "four bottlenecks" in AI development, and boldly bet that a scientific breakthrough that would change human perception would occur in 2027.
The following is the full text of the interview highlights:
01. Altman Recalls His Childhood Programming Dream, GPT - 5 Creates Snake Game in Seven Seconds
Abram: OpenAI has just released GPT - 5, and sitting next to me is Sam Altman, the CEO of OpenAI. OpenAI is reshaping the industry. Now, they are trying to build a super - intelligence that far surpasses humans in almost all fields and have just released the most powerful model to date.
We are experiencing a profound moment. Most people have never experienced such a huge technological change in their lives, but it is happening around us. In this conversation, I want to take a "time - travel" with Altman to explore the future he is trying to build, so that both you and I can truly understand what is about to happen. Hello, Altman. Thank you for participating in this interview.
Altman: Of course, I'm glad to participate.
Abram: Before we start, I want to clarify my goal. I won't ask questions about valuation, the AI talent war, or financing, as there has been a lot of coverage on these topics elsewhere.
Altman: It seems so.
Abram: The goal of our program is to explore how to use science and technology to make the future better. We believe that if people can see a better future, they can help build it. Therefore, my purpose is to time - travel with you as much as possible to different moments in the future you are trying to build and understand what it will be like.
Altman: Great, that's very good.
Abram: You recently mentioned that GPT - 4 is the "dumbest" model we use.
Abram: GPT - 4 can already outperform 90% of humans in exams such as the SAT, LSAT, and GRE, and can also pass programming exams, sommelier exams, and medical license exams. Now you've just released GPT - 5. What can GPT - 5 do that GPT - 4 can't?
Altman: First of all, an important conclusion is that you can have an AI system that performs well in these exams, but it obviously cannot replicate many of the abilities that humans are good at, which reflects the limitations of tests like the SAT.
Recall when GPT - 4 was released. If we had told you about its capabilities, you might have said, "This will have a huge impact on many jobs and people's behaviors, and perhaps there will be negative effects." But some of the positive impacts you predicted didn't fully materialize. The areas where these models excel do not fully cover what we need humans to do or value humans doing.
I think the same goes for GPT - 5. People will be shocked by its capabilities. It is very good in many aspects but still has limitations in others. People will use it to accomplish all kinds of amazing things. It will change knowledge work, learning methods, and creative ways, but human society will evolve with it, expecting more and better tools.
So, GPT - 5 is very excellent in many aspects but still has limitations in some areas. It can be called up at any time like your phone assistant or portable device, and can complete tasks you need in a few minutes or an hour, even tasks that are difficult for experts. This is truly remarkable. It has never happened in human history, and technological progress is so rapid. We are gradually getting used to it, but if we went back five or ten years and told people that this technology was coming, they might have thought it was incredible.
Abram: What features of GPT - 5 excite you the most? Which features may not meet expectations?
Altman: What excites me the most is that for the first time, I feel that I can ask the model any complex scientific or technical question and get a fairly good answer. For example, when I was in the ninth grade of junior high school, I got a TI - 83 graphing calculator and spent a long time programming a "Snake" game, which was quite popular at school. But programming for the TI - 83 was extremely laborious, time - consuming, and difficult to debug. Later, I had an idea and tried to use an early version of GPT - 5 to see if it could create a TI - 83 - style Snake game. As a result, it perfectly completed the task in 7 seconds.
I was stunned for 3 seconds, wondering if my 11 - year - old self would think it was cool or if something was missing. Then I realized that what I missed was the game itself. So I came up with an idea for a new feature and input it, and it immediately implemented it, and the game was updated in real - time. I then asked it to change the appearance and add functions. This programming experience made me feel like I was 11 years old again - being able to quickly express ideas, try new things, and play in real - time.
I used to worry that children who haven't experienced the "Stone Age" programming struggles I had would miss out on something, but now I'm excited for them because new tools can quickly turn people's ideas into reality. This is truly amazing. GPT - 5 can not only answer complex questions but also almost instantly create customized software, which is a landmark feature of the GPT - 5 era and something GPT - 4 didn't have.
Abram: I haven't used GPT - 5 much, but I've heard that it can be more deeply integrated into daily life, such as interacting with Gmail and calendars. My interaction with GPT - 4 always felt isolated. How will my relationship with GPT - 5 change?
Altman: It will be more naturally integrated into your life, connected with tools like calendars and Gmail, and will become more proactive. For example, when you wake up in the morning, it might tell you, "There were these new situations last night. I noticed that you adjusted your calendar, and I have some new ideas about the question you asked yesterday."
In the future, we will also launch consumer - grade devices. It might quietly accompany you during an interview and say to you after it's over, "The conversation just now was great, but next time you might ask me this question because I didn't answer the last one very well." It will be like a companion throughout your day.
02. Reviewing the Evolution of GPT: The Game of Predicting the Next Word Became an AI Revolution
Abram: For those who don't understand how algorithm design can improve the user experience, could you briefly summarize the current situation? Why do you find this question interesting?
Altman: We can start from history.
When developing GPT - 1, we came up with an idea that was ridiculed by experts at the time - training the model to play a "game": giving it a string of words and asking it to predict the next word. This is what's called unsupervised learning. We don't directly tell it "this is a cat" or "that is a dog", but let it infer the next possible word through word sequences. It sounds absurd, but it turns out that this way, the model can learn complex concepts such as physics, mathematics, and programming without explicit teaching. In fact, the process of human infants learning language is quite similar.
Later, we found that the performance of the model was closely related to its scale, and significant improvement required an increase across multiple orders of magnitude. The performance of GPT - 1 was indeed poor, and many experts at the time asserted that this path would not succeed. But we found the so - called "scaling law" - as computing power, memory, and data volume increased, the performance of the model would continuously improve in a predictable trend. So we firmly expanded the model in this direction and achieved unexpectedly good results.
We also introduced reinforcement learning to improve the reasoning ability by letting the model know which answers were good and which were bad. This method was initially considered too simple to bring about a qualitative breakthrough, but it turned out to be the reason for the leaps in O1, O3, and GPT - 5. Now we are exploring video models, using new data and interaction environments to further expand capabilities. It is expected that in the next few years, the progress of algorithm design will remain stable and strong.
Abram: The public may think that the upgrade from GPT - 1 to GPT - 5 is a smooth path, but it must be more complicated behind the scenes. Can you share some interesting problems you encountered before the release of GPT - 5?
Altman: Once we developed a model codenamed Orion, which was later released as GPT - 4.5. It was large - scale and had cool functions, but the user experience was not ideal. This made us realize that research not only needs to pursue "bigness" but also explore model architectures of different "shapes".
We originally always followed the scaling law, thinking that as long as the model was larger, the performance would linearly improve. But later we found that in terms of reasoning ability, there was another steeper "scaling curve", and moving along it would bring higher returns. This was a detour in research, but detours often lead to new discoveries.
In terms of the dataset, we also encountered problems. The model needs a large amount of high - quality data to learn, but sometimes it is restricted by the bottlenecks of data quality or coverage. Every day's progress is accompanied by twists and turns. For example, a certain architecture attempt may not work in the end, but the overall trend is still moving forward steadily.
Abram: GPT - 5 has been released, and you must be thinking about the future now. If I interview you one year later, what do you think you'll be thinking about?
Altman: Maybe you'll ask me, "What does it mean for AI to make new scientific discoveries? How will the world view the scientific achievements brought by GPT - 6?".
Maybe by then, these achievements won't be fully realized, but it will feel like they're just around the corner. If it really happens, the good part will be exciting, such as curing diseases; the bad part may be worrying, such as being used for biosecurity threats; and there will also be some strange new phenomena that people will get used to soon.
The speed of world change will be dizzying. The economy may grow rapidly, but humans have strong adaptability. It won't take long for people to regard these huge changes as the new normal of life.
03. AI Is Not a Tool for Laziness, but a Multiplier for the Diligent
Abram: What you mentioned made me think of the concept of "time under tension" in weightlifting. Laypeople may not know that when lifting 100 pounds, the benefits of completing it in 30 seconds are much greater than those of completing it in 3 seconds.
I think the best work I've done is inseparable from a large amount of "cognitive load time", which is crucial. Ironically, developing these tools themselves requires a huge investment of cognitive load time, but some people may think that these tools are shortcuts to avoid thinking. You might say that just as the calculator allows us to turn to solving more difficult math problems, what's the difference between the two? What's your view on this question?
Altman: This is indeed different from a calculator. Some people use ChatGPT to avoid thinking, while others use it to think more deeply. I hope the tools we design can encourage more people to use them to expand their thinking and complete more work. Society itself is a competitive arena. Theoretically, new tools may allow people to reduce their workload, but in reality, people will work harder and have higher expectations for themselves.
Just like other technologies, some people achieve more with ChatGPT, while others achieve less. But those who want to increase their cognitive load time can indeed do so through it. I've been inspired by the top 5% of the most active users. Their learning efficiency, work ability, and output are extremely amazing. I've only used GPT - 5 for a few hours and am still exploring how to interact with it. Interestingly, I just learned to use GPT - 4, and now I have to learn to use GPT - 5.
Abram: You've used GPT - 5 for some time. What's the most interesting specific task you've done with it? What tasks have impressed you the most?
Altman: The programming tasks have impressed me the most. It performs well in many aspects, but it is particularly outstanding in writing software for various needs. This means that we can express ideas in new ways, and AI can complete complex tasks. Theoretically, GPT - 4 can also answer any question, but GPT - 5 is so powerful in programming that it almost seems omnipotent. Of course, it can't handle physical - world affairs, but it can make computers complete complex work. Software is a super - powerful control tool that can achieve many functions, which is truly amazing.
In addition, the writing ability of GPT - 5 has also been greatly improved. AI writing sometimes has an annoying style, such as using a lot of dashes. Although GPT - 5 still uses dashes, many people actually like this style. Its writing quality is much better than before, but there is still room for improvement. Many people inside OpenAI say that after using GPT - 5, they feel that it is better in all indicators and has an indescribable sense of delicacy. When recalling the testing of GPT - 4, it didn't feel as good, probably because the writing of GPT - 5 is more natural and of higher quality.
04. By 2027, AI May Achieve a Generally Recognized Major Scientific Breakthrough
Abram: What you mentioned made me think of the concept of "time under tension" in weightlifting. Laypeople may not know that when lifting 100 pounds, the benefits of completing it in 30 seconds are much greater than